CN107609535A - Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model - Google Patents
Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model Download PDFInfo
- Publication number
- CN107609535A CN107609535A CN201710901019.7A CN201710901019A CN107609535A CN 107609535 A CN107609535 A CN 107609535A CN 201710901019 A CN201710901019 A CN 201710901019A CN 107609535 A CN107609535 A CN 107609535A
- Authority
- CN
- China
- Prior art keywords
- mrow
- tree
- face
- shared pool
- hybrid coordination
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Abstract
A kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and localization method, the described method comprises the following steps;It is modeled each face's calibration point as a part for global feature point, acquisition is based on shared pool hybrid coordination tree model, and the model includes:Department pattern, shape and the composite character vector of tree construction;Face's calibration point of various pieces under different visual angles is mixed into general characteristic point, for representing the mark point under different visual angles;The compound tree allowed under different visual angles shares the department pattern of tree construction, and view is modeled with low complex degree;By based on all parameters of shared pool hybrid coordination tree model, making a distinction property is trained in restriction range.The present invention is modeled each facial calibration point as a part for global feature point, and captures the change in topology caused by visual angle using overall situation mixing.
Description
Technical field
The present invention relates to Face datection, Attitude estimation and positioning field, more particularly to one kind to be based on shared pool compound tree mould
Face datection, Attitude estimation and the localization method of type.
Background technology
Human face analysis is the basic task of computer vision field.Although length is achieved in terms of Face datection in the industry
The progress of foot, but the reliable prediction marked to head pose and face is still challenging, particularly unconfined " true
In real environment " image.Although current learned Face datection precision is up to 99%, in fact, in extreme circumstances, by
In the reason for the visual angle change and elastic deformation, this is still challenging.
Face datection by distinguish training scanning window grader control, at present using it is most common be Viola Jones
Detector[1], because algorithm is in OpenCV[2]In realize, it is therefore very easy to use.Pose estimation is often in video field
It is resolved under scape or controlled experiment environment, now assumes that solved the problems, such as Face datection, such as MultiPIE[3]Or
FERE benchmark[4].Most methods use explicit 3D models or the model based on 2D views at present.Face's demarcation estimation can trace
To movable appearance model (AAMs) and the classical way of elastic graph matching.Prior art concentrate on establish local part detector it
On global space model, sometimes referred to as local restriction model (CLMs).It is worth noting that, all these work all assume
The spatial model of one intensive connection is, it is necessary to substantially matching algorithm.In current computer vision field, prior art does not have
Solves the task of Face datection, pose estimation and demarcation estimation simultaneously.
Face datection, Attitude estimation and mark positioning are considered as different technical problems all the time, such as scan window
Mouth grader, the eigenspace method based on view and elastic graph model.
It is worth noting that, it is most of in the past on marking the method estimated all to use the elastic graph of intensive connection, it is this
Method is difficult to optimize.Therefore, many improvement directions in the field all lay particular emphasis on optimized algorithm at present, to avoid the occurrence of Local Minimum
Value.
The content of the invention
The invention provides a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and localization method, sheet
Invention is modeled each facial calibration point as a part for global feature point, and using the overall situation mixing come capture due to regarding
Change in topology caused by angle, it is described below:
A kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and localization method, methods described include with
Lower step;
It is modeled each face's calibration point as a part for global feature point, acquisition is based on shared pool compound tree mould
Type, the model include:Department pattern, shape and the composite character vector of tree construction;
Face's calibration point of various pieces under different visual angles is mixed into general characteristic point, for representing under different visual angles
Mark point;
The compound tree allowed under different visual angles shares the department pattern of tree construction, and view is modeled with low complex degree;
By based on all parameters of shared pool hybrid coordination tree model, making a distinction property is trained in restriction range.
It is described to be specially based on shared pool hybrid coordination tree model:
S (I, L, m)=Appm(I,L)+Shapem(L)+αm
Wherein, S (I, L, m) is the department pattern of tree construction;Appm(I, L) is composite character vector;Shapem(L) it is shape
Shape model;αmIt is and scalar deviation or priori value that compound tree m is associated;wiFor i-th section template;φ(I,li) be characterized to
Amount;(aij,bij,cij,dij) it is constrained parameters;I and j represents any two parts pixel.
Methods described also includes:Optimization to the shape, it is specially:
Shapem(L)=- (L- μm)TΛm(L-μm)+constant
Wherein, (μ, Λ) is constrained parameters (a, b, c, d) Reparameterization;ΛmIt is the sparse concentration matrix of block, nonzero term
Corresponding to i;J corresponds to Em;μmFor ideal form model;Constant is constant.
The compound tree allowed under different visual angles shares the department pattern of tree construction, and view is modeled with low complex degree
Specially:By the non-individual body between common view angle and extreme visual angle, it is written as
Wherein, f (m) is from 1 to M by blended index, is mapped to function of the smaller template index from 1 to M';
For M':As M'=M, any numerical value is not shared;As M'=1, numerical value is shared between all views.
The beneficial effect of technical scheme provided by the invention is:
1st, the present invention has used multi views tree in the work of mark estimation, can be carried out by Dynamic Programming global excellent
Change, this is huge progress compared with prior art;
2nd, the present invention shows in all tasks must be better than past achievement, particularly under extreme observation visual angle,
Compared with prior art, this method can capture sizable global elastic construction, be able to observe that more mark points.
Brief description of the drawings
Fig. 1 is the flow chart of a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and localization method;
Fig. 2 is the compound tree sample schematic diagram under various visual angles;
Fig. 3 a are comparing result schematic diagram of a variety of methods in Face datection;
Fig. 3 b are that a variety of methods compare result schematic diagram in another pair of Face datection;
Fig. 4 a are the accumulated error distribution curve schematic diagram on various visual angles AAMs data sets;
Fig. 4 b are face.com[5]Accumulated error distribution curve schematic diagram on data set;
Fig. 5 a be this method on MultiPIE data sets for the mark positioning performance schematic diagram of positive face;
Fig. 5 b are the error flag schematic diagram based on the positioning of shared pool hybrid coordination tree model;
Wherein, the standard deviation schematic diagram of each ellipse representation position error.
Fig. 6 a are average localization error schematic diagram of a variety of methods on MultiPIE data sets;
Fig. 6 b are average localization error schematic diagram of a variety of methods on AFW data sets.
Embodiment
To make the object, technical solutions and advantages of the present invention clearer, embodiment of the present invention is made below further
It is described in detail on ground.
In order to solve problem present in background technology, the embodiment of the present invention proposes one kind and is based on shared pool compound tree mould
Face datection, Attitude estimation and the localization method of type, for Face datection, Attitude estimation and position these three tasks for, this
Method improves current highest level, and the amplitude improved is sizable.The embodiment of the present invention thinks a unified side
Method can be such that problem becomes easier to a bit.For example, many work carried out to face calibration, can assume that image is examined by face first
Device pre-filtering is surveyed, therefore nearby specified deviation can be caused.
Show that multi views tree is an effective alternative solution by research at present, reason is as follows:
1) global optimization can be carried out by Dynamic Programming;2) sizable global elastic construction can still be captured.
Research shows:By the hybrid coordination tree model based on shared pool, can be found using efficient dynamic programming algorithm
Globally optimal solution, and sizable global elastic construction can be captured.
Embodiment 1
The embodiment of the present invention proposes a kind of method based on partial sharing pond hybrid coordination tree model, referring to Fig. 1, this method bag
Include following steps:
101:It is modeled each face's calibration point as a part for global feature point, obtains and mixed based on shared pool
Tree-model, above-mentioned model include:Department pattern, shape and the composite character vector of tree construction;
102:Face's calibration point of various pieces under different visual angles is mixed into general characteristic point, for representing different visual angles
Under mark point;
103:The compound tree allowed under different visual angles shares the department pattern of tree construction, and view is built with low complex degree
Mould;
104:By based on all parameters of shared pool hybrid coordination tree model, making a distinction property is trained in restriction range.
In summary, the embodiment of the present invention regard each facial calibration point as entirety by above-mentioned steps 101- steps 104
A part for characteristic point is modeled, and captures the change in topology caused by visual angle using overall situation mixing, by based on
The hybrid coordination tree model of shared pool, globally optimal solution is found using efficient dynamic programming algorithm, and can captured quite big
Global elastic construction
Embodiment 2
The scheme in embodiment 1 is further introduced with reference to specific calculation formula, accompanying drawing, example, in detail
See below description:
201:Being modeled using each face's calibration point as a part for global feature point (will each face's calibration point
Regard a part for tree construction as), and capture the change in topology as caused by visual angle using overall situation mixing;
Wherein, refer to face's calibration point of various pieces under different visual angles being mixed into overall spy referring to Fig. 2, overall situation mixing
Sign point.
The department pattern of tree construction:T will each be setm=(Vm;Em) write as the tree of a linear parameterization[6], its
In, m represent mixing andVmFor the division center of each tree;EmFor the marginal texture of each tree;V is all VmCollection
Close;M value is positive integer.
S (I, L, m)=Appm(I,L)+Shapem(L)+αm (1)
Image, and l are represented with an Ii=(xi;yi) be used to represent the location of pixels of each tree construction i-th section, obtain
To the configuration L={ l of each parti:i∈V}:
Wherein, xi;yiThe respectively transverse and longitudinal coordinate of i-th section pixel;S (I, L, m) is the department pattern of tree construction; Appm
(I, L) is composite character vector;Shapem(L) it is shape;αmIt is and scalar deviation or " priori that compound tree m is associated
Value ";wiFor i-th section template;φ(I,li) it is characterized vector;(aij,bij,cij,dij) it is constrained parameters;I and j represents any
Two parts pixel.
Formula (2) is represented the l from image IiThe characteristic vector (for example, HoG descriptions) for extracting to obtain in place's pixel is write
Enter Φ (I, li)。
Mixing particular space arrangement of the formula (3) to L portion is scored, wherein dx=xi-xjAnd dy=yi-yjIt is i-th
Partly relative to the displacement of jth part.
Each single item in formula (3) can be regarded as the space constraint introduced between any two parts (i.e. i and j),
(aij,bij,cij,dij) resting position and rigidity of each constraint is determined.
Because the location variable l in formula (2)iCan only occur with linear and quadratic term, so shape can be rewritten
For:
Shapem(L)=- (L- μm)TΛm(L-μm)+constant(4)
Wherein, (μ, Λ) is constrained parameters (a, b, c, d) Reparameterization;The norm of Gauss is similarly to joining naturally
Numberization.Now, ΛmIt is the sparse concentration matrix of block, nonzero term corresponds to i;J corresponds to Em;μmFor ideal form model;
Constant is constant.
And if only if, and quadratic constraints item a and c are when bearing[7], ΛmFor just half fixed number.This corresponds to from ideal form μmDeformation
For the punishment score of L configuration.The Λ associated with minimal eigenvaluemCharacteristic vector represent the deformation associated with small punishment
Model.
202:The compound tree allowed under different visual angles shares department pattern (i.e. S (I, L, m)), and with low complex degree to view
It is modeled;
Above-mentioned formula (1) needs each single models of compound tree m of corresponding i-th sectionHowever, the change at part visual angle
Change appears likely to be consistent.In the case of extreme visual angle (such as from the nearly zero angle in side, corresponding model is only
Formwork erection type), independent model is by all modelsIn use single hybrid coordination tree model for marginal portion.
The non-individual body that this method have studied between common view angle (- 60 °~60 °) and extreme visual angle both of these case is (i.e. general
Shared content between intervisibility angle and extreme visual angle), it is written asWherein f (m) is to reflect blended index (from 1 to M)
It is mapped to the function of smaller template index (from 1 to M').For M':As M'=M, any numerical value is not shared;Work as M'=1
When, numerical value is shared between all views.
203:All parameters of the hybrid coordination tree model based on shared pool (including:S (I, L, m), ShapemAnd App (L)m
(I, L)) making a distinction property is trained in restriction range.
Learn this hybrid coordination tree model based on shared pool, assume initially that a scene supervised completely, this method carries
The positive sample image with mark and hybrid tag, and the negative sample image without face are supplied.
Hybrid coordination tree model based on shared pool distinguishes learning model and apparent parameter using structuring prediction framework.Need first
Estimate the marginal texture E of each compound treem.Although it is natural to set the modeling for human body[8,6,9], but the day of face's demarcation
Right tree construction is unclear.Assuming that mark meets Gaussian Profile, then using Chow-Liu algorithms[10]Best it is explained to find
Determine the maximum likelihood tree construction of the mark position of compound tree.
Positive example { the I of given markn,Ln,mnAnd negative example { In, it now will define one and be similar in bibliography [6]
The structuring prediction object function of proposition.Z is made firstn={ Ln,mn}.Pay attention to, equation (1) is in segment template ω, constrained parameters
(aij,bij,cij,dij) and mix deviations α in be linear.These parameters are connected into single vector β, now can by S (I,
L, m) it is written as:
S (I, z)=β Φ (I, z) (5)
Vectorial Φ (I, z) is sparse, nonzero term be present in the monospace corresponding to compound tree m.
Learn another form of model now, globally optimal solution S (I, z) tried to achieve by the model:
Wherein, C is adjustment parameter;Pos is positive sample;Neg is negative sample;Φ(In,zn) it is structuring anticipation function;βk
For parameter vector;ξnFor slack variable;HereinScore is represented, by the minimum β of score and meets constraint bar
Φ (the I of partn,zn) obtain globally optimal solution S (I, z).
Above-mentioned constraint qualification, positive example should be better than 1 (boundary value), and for the negative of portion and all configurations of compound tree
Face example should be less than -1.Object function uses slack variable ξnThe behavior of these constraints is violated in punishment.
This method writes out quadratic constraints item (a in parameter vector βij,cij) index K.Related beam of breaking a promise ensures shape
Matrix Λ is just half fixed number.Then use[6]In double coordination solvers solve this problem, it receives passive constraint.
In summary, the embodiment of the present invention is modeled each facial calibration point as a part for global feature point,
And the change in topology caused by visual angle is captured using overall situation mixing, by the hybrid coordination tree model based on shared pool, use
Efficient dynamic programming algorithm (i.e. formula (6)) finds globally optimal solution, and can capture sizable global elasticity and tie
Structure (is used for solving elastic deformation in Face datection, that is, face caused by expression a variety of causes because deform).
Embodiment 3
Feasibility checking is carried out to the scheme in Examples 1 and 2 with reference to specific accompanying drawing, example, it is as detailed below to retouch
State:
900 positive examples of the database that this experiment uses from MultiPIE and INRIAPerson databases [11] (its is past
Toward being the Outdoor Scene not comprising people) 1218 negative examples.Each mark defined in MultiPIE is considered as a part.It is all
Visual angle shares 99 parts.5 × 5 HoG cells are each partly represented as, spatial volume size is 4.13 are used at positive visual angle
Individual visual angle and 5 expression formulas, common property give birth to 18 hybrid representations.For simplicity, this method is not forced between left/right view
Symmetry.
In addition, this method is also labelled with true environment (AFW) test set:Shared pool is based in order to further assess this
Hybrid coordination tree model, randomly select image from Flickr image sets, but ensure that each image comprises at least a clearly face
Portion.It finally obtained the data set of 205 image for possessing 468 faces.Image includes rambling background mostly,
Varied widely at face visual angle and outward appearance (age, glasses, cosmetic, the colour of skin, expression etc.) aspect.Each face marked
Frame, along pitching and yaw direction indicate 6 marks (eye center, nose, two corners and center) and discrete visual angle (- 90 °~
90 °, every 15 ° one) and (left, middle and right) visual angle.This data set with[12,13,14,15]In approximation " true environment " collect more
Difference is shown on non-frontal face in individual annotation and single image.
Due to there are three kinds of detections, therefore three kinds of different evaluation criterias are employed to the hybrid coordination tree model based on shared pool,
Corresponding to Face datection, Attitude estimation and mark positioning.
For Face datection, quasi- full curve (Precision-Recall) is looked into weigh the detection of this model using looking into
Energy.It is one of important indicator of Performance Evaluation of three-dimensional body retrieval to look into quasi- full curve of looking into.Recall is tried to achieve according to below equation
And Precision, make looking into and quasi- look into full curve:
Wherein, Recall is recall ratio, NzIt is the quantity of correct retrieval object, NrIt is the quantity of all related objects.
Wherein, Precision is precision ratio, NallIt is the quantity of all retrieval objects.
For Attitude estimation, using deviation accumulation distribution curve (cumulative error distribution
Curves) performance is estimated weigh this model.Deviation accumulation distribution curve can completely describe the general of real number stochastic variable x
Rate is distributed, and is the integration of probability density function.
Positioned for mark, this model is weighed using normalization pixel error (normalize pixel error)
Estimate performance.Normalization is a kind of mode of simplified calculating, will there is the expression formula of dimension, by conversion, is turned to nondimensional
Expression formula, turn into scale.
Normalization is for the convenience of subsequent data processing, next to that ensureing to accelerate convergence during program operation first.Normalizing
The specific effect changed is to conclude the statistical distribution of unified samples.
For Face datection, this hybrid coordination tree model based on shared pool will be contrasted with following five kinds of methods:
(1) the positive faces of the OpenCV based on Viola-Jones+side face detector;
(2) positive face+side face detector of enhancing[16];
(3) deformable part sub-model (DPM)[17,18](use and model identical data);
(4) Google Picasa human-face detectors[19], by checking scoring manually;
(5) face.com face detectors, report detection, visual angle and mark.
For Attitude estimation, this hybrid coordination tree model based on shared pool will be contrasted with following two methods:
(1) various visual angles AAMs models:Use[20]In code train AAM for each visual angle, and shown in test image
Go out the particular figure model with minimal reconstruction error;
(2) face.com face detectors.
Positioned for mark, this hybrid coordination tree model based on shared pool will be contrasted with following four method:
(1) various visual angles AAMs models;
(2) local restriction model (CLMs):It is used herein as[21]In existing code.This operational representation MultiPIE
Current state-of-the-art mark estimated result;
(3)face.com;
(4) Oxford face markers detector[22]。
Experimental result
First, Face datection:
The result of all method for detecting human face is summarized in fig.4.This hybrid coordination tree model based on shared pool is substantially better than only
Have 2 visual angles Viola Jones detectors and[19]In detector, be just slightly below Google Picasa and face.com
Human-face detector.The model is adjusted for big face, therefore it is high-visible to mark.This method is to all algorithms
(including benchmark) has carried out other assessment, and the sample that 150 pixels are more than for height (shares 329, or 70% in AFW
Face).In this case, this model inspection result and Google Picasa and face.com are suitable (Fig. 4 b).It is and existing
Nowadays high-definition picture is fairly common for HD video and mega pixel camera.
It can see by Fig. 4 b, rigid various visual angles HoG benchmark are better than now widely used human-face detector, reach
77.4% mean accuracy (AP).Potential hub-and-spoke configuration part is added, it is supervised, adds structure sexual intercourse, these
Method each contributes to improving performance, and AP has respectively reached 87.8%, 88.6% and 92.9%.
2nd, Attitude estimation:
Fig. 5 (a) and 5 (b) show the accumulated error distribution curve in two datasets.The posture of displaying estimation herein is at certain
Face's period in individual error tolerance.The independent model effect is best, and when MultiPIE allows ± 15 ° of errors, detection is accurate
True rate has reached 99.9%.From assuming that the MultiPIE of detection is different (image is centered on face), this sentences more real side
Formula assesses the performance of AFW data sets:This method assesses the face result that finds of given algorithm, and by the mistake in Attitude estimation
Detection is calculated as unlimited mistake.Because AAMs models give it and used on AFW test sets here without the detector of correlation
What actual value bounding box was carried out most preferably may initialization (being represented in figure 5b with *).
Performance of all curves in AFW is all declining (showing that the difficulty of data set is larger), particularly various visual angles AAMs,
This shows fitness deficiencies of the AAMs to new data.The independent model shows again optimum performance, correctly by 81.0%
Face calibration in ± 15 ° of error ranges.
In general, this model is similar with the performance of Multiview-HoG/Star benchmark, and is greatly better than
Face.com and various visual angles AAMs.It note that this method will not be punished the false positive of Attitude estimation;If by false positive
Punish as incorrect Attitude estimation (because they are worse detectors), the star-like benchmark performance meetings of various visual angles HoG/ herein
It is even worse.In view of the difficulty of this no bound data is big, therefore the result impressive of this method.
3rd, mark positioning:
MultiPIE is firstly evaluated in Fig. 6 a herein for the performance of positive face.All benchmark all show well, but the base
In shared pool hybrid coordination tree model (mean error is the relative error of 4.39 pixels/2.3%) still better than state-of-the-art CLM moulds
Type[21](4.75 pixel/2.8%).When assessing all visual angles, it can be seen that the efficiency of most of benchmark is all under
Drop, particularly CLMs (Fig. 6 a).Significantly, since CLM and Oxford processing is approximate positive face data, so here
Only the face between p- 45 ° and 45 ° is assessed, wherein the mark of all positive faces is high-visible (marked in Fig. 6 a
For *).Even if other models have this advantage, the effect for being somebody's turn to do the hybrid coordination tree model based on shared pool also greatly exceed all bases
It is accurate.
In AFW (Fig. 6 b), missing inspection is calculated as unlimited position error by this method again.As a result it is clear to be directly displayed at mark
(including 329 face examples in AFW test sets) in visible face.The hybrid coordination tree model based on shared pool is another
Secondary to have reached best effect, 76.7% face mark position error is below face size 5%.AAMs and CLM's standard
True property is still declining, and this shows the image that these popular methods can not be generalized under true environment well.Finally, the base
Huge spread between the hybrid coordination tree model of shared pool and star-like benchmark shows that the tree construction that this method proposes can obtain really
Useful elastic construction.
It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention
Sequence number is for illustration only, does not represent the quality of embodiment.
The foregoing is only presently preferred embodiments of the present invention, be not intended to limit the invention, it is all the present invention spirit and
Within principle, any modification, equivalent substitution and improvements made etc., it should be included in the scope of the protection.
Bibliography:
[1]P.Viola and M.J.Jones.Robust real-time face detection.IJCV,2004.2
[2] OpenCV is a cross-platform computer vision library based on BSD licenses (increasing income) distribution.
[3]R.Gross,I.Matthews,J.Cohn,T.Kanade,and S.Baker.Multi-pie.Image and
Vision Computing,2010.2,4
[4]P.Phillips,H.Moon,S.Rizvi,and P.Rauss.The feret evaluation
methodology for face-recognition algorithms.IEEE TPAMI,2000.2
[5]http://face.com/.2
[6]Y.Yang and D.Ramanan.Articulated pose estimation using flexible
mixtures of parts.In CVPR 2011,2011.2,4
[7]H.Rue and L.Held.Gaussian Markov random fields:theory and
applications,volume 104. Chapman&Hall,2005.3
[8]P.F.Felzenszwalb and D.P.Huttenlocher.Pictorial structures for
object recognition.IJCV, 2005.3,4
[9]S.Ioffe and D.Forsyth.Mixtures of trees for object recognition.In
CVPR 2001.2,4
[10]C.Chow and C.Liu.Approximating discrete probability distributions
with dependence trees. IEEE TIT,1968.4
[11]N.Dalal and B.Triggs.Histograms of oriented gradients for human
detection.In CVPR 2005.4
[12]V.Jain and E.Learned-Miller.Fddb:A benchmark for face detection
in unconstrained settings.Technical Report UMCS-2010-009,University of
Massachusetts,Amherst,2010.4
[13]http://vis-www.cs.umass.edu/lfw/results.html.2,4
[14]M.Kostinger,P.Wohlhart,P.M.Roth,and H.Bischof.An-¨notated facial
landmarks in the wild:A large-scale,realworld database for facial landmark
localization.In First IEEE International Workshop on Benchmarking Facial
Image Analysis Technologies 2011.4
[15]P.N.Belhumeur,D.W.Jacobs,D.J.Kriegman,and N.Kumar.Localizing
parts of faces using a consensus of exemplars.In CVPR 2011.2,4,6
[16]Z.Kalal,J.Matas,and K.Mikolajczyk.Weighted sampling for large-
scale boosting.In BMVC 2008.2,5
[17]P.Felzenszwalb,R.Girshick,D.McAllester,and D.Ramanan.Object
detection with discriminatively trained partbased models.IEEE TPAMI,2009.2,3,5
[18]http://www.cs.brown.edu/~pff/latent/voc-release4.tgz.5
[19]http://picasa.google.com/.2
[20]D.-J.Kroon.Active shape model and active appearance model.http://
www.mathworks.com/ matlabcentral/fileexchange/26706.5
[21]J.Saragih,S.Lucey,and J.Cohn.Deformable model fitting by
regularized landmark mean-shift.IJCV,2011.2,3,6
[22]M.Everingham,J.Sivic,and A.Zisserman.“Hello!My name is...Buffy”–
automatic naming of characters in TV video.In BMVC 2006.2,6。
Claims (4)
1. a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and localization method, it is characterised in that the side
Method comprises the following steps;
It is modeled each face's calibration point as a part for global feature point, acquisition is based on shared pool hybrid coordination tree model,
The model includes:Department pattern, shape and the composite character vector of tree construction;
Face's calibration point of various pieces under different visual angles is mixed into general characteristic point, for representing the mark under different visual angles
Point;
The compound tree allowed under different visual angles shares the department pattern of tree construction, and view is modeled with low complex degree;
By based on all parameters of shared pool hybrid coordination tree model, making a distinction property is trained in restriction range.
2. a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and positioning side according to claim 1
Method, it is characterised in that described to be specially based on shared pool hybrid coordination tree model:
S (I, L, m)=Appm(I,L)+Shapem(L)+αm
<mrow>
<msub>
<mi>App</mi>
<mi>m</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>I</mi>
<mo>,</mo>
<mi>L</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>&Element;</mo>
<msub>
<mi>V</mi>
<mi>m</mi>
</msub>
</mrow>
</munder>
<msubsup>
<mi>w</mi>
<mi>i</mi>
<mi>m</mi>
</msubsup>
<mo>&CenterDot;</mo>
<mi>&phi;</mi>
<mrow>
<mo>(</mo>
<mi>I</mi>
<mo>,</mo>
<msub>
<mi>l</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<msub>
<mi>Shape</mi>
<mi>m</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>L</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mi>j</mi>
<mo>&Element;</mo>
<msub>
<mi>E</mi>
<mi>m</mi>
</msub>
</mrow>
</munder>
<msubsup>
<mi>a</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mi>m</mi>
</msubsup>
<msup>
<mi>dx</mi>
<mn>2</mn>
</msup>
<mo>+</mo>
<msubsup>
<mi>b</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mi>m</mi>
</msubsup>
<mi>d</mi>
<mi>x</mi>
<mo>+</mo>
<msubsup>
<mi>c</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mi>m</mi>
</msubsup>
<msup>
<mi>dy</mi>
<mn>2</mn>
</msup>
<mo>+</mo>
<msubsup>
<mi>d</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mi>m</mi>
</msubsup>
<mi>d</mi>
<mi>y</mi>
</mrow>
Wherein, S (I, L, m) is the department pattern of tree construction;Appm(I, L) is composite character vector;Shapem(L) it is shape mould
Type;αmIt is and scalar deviation or priori value that compound tree m is associated;wiFor i-th section template;φ(I,li) it is characterized vector;
(aij,bij,cij,dij) it is constrained parameters;I and j represents any two parts pixel.
3. a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and positioning side according to claim 2
Method, it is characterised in that methods described also includes:Optimization to the shape, it is specially:
Shapem(L)=- (L- μm)TΛm(L-μm)+constant
Wherein, (μ, Λ) is constrained parameters (a, b, c, d) Reparameterization;ΛmIt is the sparse concentration matrix of block, nonzero term is corresponding
In i;J corresponds to Em;μmFor ideal form model;Constant is constant.
4. a kind of Face datection based on shared pool hybrid coordination tree model, Attitude estimation and positioning side according to claim 1
Method, it is characterised in that the compound tree allowed under different visual angles shares the department pattern of tree construction, is entered with low complex degree to view
Row modeling is specially:
By the non-individual body between common view angle and extreme visual angle, it is written as
Wherein, f (m) is from 1 to M by blended index, is mapped to function of the smaller template index from 1 to M';
For M':As M'=M, any numerical value is not shared;As M'=1, numerical value is shared between all views.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710901019.7A CN107609535A (en) | 2017-09-28 | 2017-09-28 | Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710901019.7A CN107609535A (en) | 2017-09-28 | 2017-09-28 | Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107609535A true CN107609535A (en) | 2018-01-19 |
Family
ID=61059344
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710901019.7A Pending CN107609535A (en) | 2017-09-28 | 2017-09-28 | Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107609535A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109949368A (en) * | 2019-03-14 | 2019-06-28 | 郑州大学 | A kind of human body three-dimensional Attitude estimation method based on image retrieval |
CN110210331A (en) * | 2019-05-14 | 2019-09-06 | 安徽大学 | A kind of estimation method of human posture of combination tree-model and Star Model |
CN110728222A (en) * | 2019-09-30 | 2020-01-24 | 清华大学深圳国际研究生院 | Pose estimation method for target object in mechanical arm grabbing system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060088207A1 (en) * | 2004-10-22 | 2006-04-27 | Henry Schneiderman | Object recognizer and detector for two-dimensional images using bayesian network based classifier |
CN102654903A (en) * | 2011-03-04 | 2012-09-05 | 井维兰 | Face comparison method |
CN103473529A (en) * | 2013-08-26 | 2013-12-25 | 昆明学院 | Method and device for recognizing faces through multi-angle imaging |
CN103514432A (en) * | 2012-06-25 | 2014-01-15 | 诺基亚公司 | Method, device and computer program product for extracting facial features |
CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
-
2017
- 2017-09-28 CN CN201710901019.7A patent/CN107609535A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060088207A1 (en) * | 2004-10-22 | 2006-04-27 | Henry Schneiderman | Object recognizer and detector for two-dimensional images using bayesian network based classifier |
CN102654903A (en) * | 2011-03-04 | 2012-09-05 | 井维兰 | Face comparison method |
CN103514432A (en) * | 2012-06-25 | 2014-01-15 | 诺基亚公司 | Method, device and computer program product for extracting facial features |
CN103473529A (en) * | 2013-08-26 | 2013-12-25 | 昆明学院 | Method and device for recognizing faces through multi-angle imaging |
CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
Non-Patent Citations (1)
Title |
---|
赵静: "基于树结构的人脸属性识别", 《万方学位论文数据库》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109949368A (en) * | 2019-03-14 | 2019-06-28 | 郑州大学 | A kind of human body three-dimensional Attitude estimation method based on image retrieval |
CN110210331A (en) * | 2019-05-14 | 2019-09-06 | 安徽大学 | A kind of estimation method of human posture of combination tree-model and Star Model |
CN110728222A (en) * | 2019-09-30 | 2020-01-24 | 清华大学深圳国际研究生院 | Pose estimation method for target object in mechanical arm grabbing system |
CN110728222B (en) * | 2019-09-30 | 2022-03-25 | 清华大学深圳国际研究生院 | Pose estimation method for target object in mechanical arm grabbing system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhu et al. | Face detection, pose estimation, and landmark localization in the wild | |
Yu et al. | Pose-free facial landmark fitting via optimized part mixtures and cascaded deformable shape model | |
CN111695522B (en) | In-plane rotation invariant face detection method and device and storage medium | |
Oliva et al. | Scene-centered description from spatial envelope properties | |
Maji et al. | Action recognition from a distributed representation of pose and appearance | |
Pons-Moll et al. | Posebits for monocular human pose estimation | |
Smith et al. | Recovering facial shape using a statistical model of surface normal direction | |
Jogan et al. | Robust localization using an omnidirectional appearance-based subspace model of environment | |
Zhao et al. | Rotation invariant point cloud analysis: Where local geometry meets global topology | |
CN107609535A (en) | Face datection, Attitude estimation and localization method based on shared pool hybrid coordination tree model | |
Dong | Optimal Visual Representation Engineering and Learning for Computer Vision | |
CN104050628A (en) | Image processing method and image processing device | |
WO2021248908A1 (en) | Gaussian process regression-based ground extraction method for three-dimensional point cloud of outdoor scene | |
Shah et al. | A novel 3D vorticity based approach for automatic registration of low resolution range images | |
CN105488491A (en) | Human body sleep posture detection method based on pyramid matching histogram intersection kernel | |
Chen et al. | Inferring unseen views of people | |
Hu et al. | Human interaction recognition using spatial-temporal salient feature | |
Bhattacharya et al. | Statistics on manifolds and landmarks based image analysis: A nonparametric theory with applications | |
Gattone et al. | A shape distance based on the Fisher–Rao metric and its application for shapes clustering | |
Koo et al. | Recovering the 3D shape and poses of face images based on the similarity transform | |
Tamminen et al. | Sequential Monte Carlo for Bayesian matching of objects with occlusions | |
Liu et al. | Human-centered saliency detection | |
Liu et al. | Mental health diagnosis of college students based on facial recognition and neural network | |
Fang et al. | MR-CapsNet: a deep learning algorithm for image-based head pose estimation on CapsNet | |
Wu et al. | Towards fitting a 3D dense facial model to a 2D image: A landmark-free approach |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180119 |